Hiring & employment screening AI

An internal team at Amazon built an experimental recruiting engine — ~500 models scoring resumes one to five stars — trained on ten years of the company's own hiring, a period whose hires were predominantly male. The models learned that history: they penalized the word 'women's' and downgraded women's-college graduates. The team found the bias, patched the named terms, and concluded the patch could not guarantee neutrality against unknown proxies — the model had learned the pattern, not the words — and the project was scrapped. The corpus's cleanest case of an organization detecting its own model's bias and choosing abandonment.

Workday, an applicant-tracking platform whose AI screening features operate inside thousands of employers' hiring pipelines at once, under a live federal collective action testing whether the vendor is directly liable as the employers' agent. On the litigation record the court sustained the agent theory at dismissal (2024) and preliminarily certified a nationwide age-discrimination collective (2025); a 2026 discovery ruling held the vendor's bias-testing privileged — the testing exists and is legally unreachable. Nothing here is a finding of discrimination: these are allegations and procedural rulings, not a verdict.

Unilever's graduate-hiring pipeline chained a games-based assessment with HireVue's automated video-interview scoring; the deployer reports ~90% faster time-to-hire, ~50,000 interview hours saved, ~£1M annual savings, and a 16% diversity improvement — every figure company- or vendor-reported, none independently audited. What makes it the domain's two-sided case is that both vendors' audit machinery is public in its honest but partial form: a cooperative academic audit of the games vendor, and the video vendor retiring its facial-analysis input under scrutiny. The blind spot: rejected candidates never re-enter the outcome data, so the effects are measured on hires only.

Amazon times every task in its fulfilment centres and compares each eligible worker's week against the other workers doing the same job at the same site; the slowest five percent may be disciplined on a ladder that ends in termination. A letter from an attorney for Amazon to a federal labour board described a system that 'automatically generates any warnings or terminations regarding quality or productivity without input from supervisors' and named about 300 workers terminated for productivity at one Baltimore fulfilment centre in thirteen months. Amazon's answer, then and since, is that the terminations are not automatic and that it has no quotas at all — 'individual performance is evaluated ... in relation to how the entire site's team is performing.' Both statements describe the same design. Because the threshold is a share of the site rather than a number, there is nothing fixed to disclose, which is simultaneously the employer's legal defence and two state regulators' finding of violation. What the law built in response is a read right: tell the worker the rule, and give them ninety days of their own speed data. Not a limit on the rule.

Amazon Flex rates its contract drivers into four standing tiers from app telemetry, gates the offer of work on the tier, and ends the account at a threshold by automated email. What makes it an atlas case is the shape of the correction channel rather than the shape of the score: the act is instant and complete, the reasons are thin, the appeal is ten days by email answered in weeks, and the only escalation beyond email carried a filing fee larger than two blocks of net pay. Then, in 2025, one city legislated the correction link back open — notice before the act, the reasons and the specific incidents, all the records relied on, an investigation before the account goes dark, and a challenge that costs nothing — and a federal appeals court left the ordinance standing. The same scorer now runs under two different correction regimes at once, and the scoring rule itself has never been reviewed by anybody, anywhere.

A pre-hire assessment vendor administering, on its own marketing, more than 30 million assessments a year across 90 countries sells three instruments into other companies' hiring pipelines: a computer-adaptive forced-choice personality test, a gamified working-memory test, and an asynchronous video product that scores machine transcripts against the personality test's own constructs. What makes this the atlas's unusual hiring case is where the dispute sits. No model here is alleged to have misbehaved. The contested link is between two sets of documents the same company wrote — buyer-facing technical manuals reporting what its own studies measured, and sales pages saying the tools are fair for all and scientifically proven not to have bias. A civil-liberties organisation read the first against the second and took the gap to two federal agencies at once, deception to one and discrimination to the other. Twenty-seven months later the manuals have been pulled from public view, the marketing is unchanged, all three products are still sold, and neither agency has publicly moved. Nothing in this record has been found by any regulator, court or independent auditor, and Aon disputes the claims.

Checkr's screening layer sits one company upstream of every gig platform that uses it, and it never deactivates anybody. It retrieves criminal records, matches them to a named applicant against an identity graph it markets as covering 96 percent of United States adults, normalises the charge data, applies each platform customer's own eligibility matrix, and furnishes a report; the platform then fires its own access decision. Three things make this record unlike the others in its domain. The failing component is a shared upstream vendor, so one wrong item is not one wrong decision but a file that travels to every platform sharing it. The correction channel is compulsory rather than discretionary — federal law gives a consumer a reinvestigation within thirty days — and the pleaded adverse action ran the day after the report. And the federal regulator wrote the operating rules for exactly this kind of pipeline between 2021 and 2024, then withdrew all four of them on 12 May 2025 while the statutes underneath stayed in force. Nothing here has been adjudicated against the vendor, and no regulator has taken public action against it.

A statute written against the polygraph in 1959 reached an AI hiring screen in 2024, and what it reached it by was a missing sentence. Massachusetts bans lie detector tests in employment and has required since 1985 that every employment application carry a one-sentence notice saying so; the provision then went roughly forty years unenforced. A CVS applicant alleged his application carried no notice and that the HireVue video interview he sat — uploaded, as pleaded, to Affectiva for affect analysis and scored into an employability score — was a lie detector test within the statute's definition. In February 2024 Judge Patti B. Saris denied both of CVS's motions in full: the private right of action reaches the notice provision, and being denied information he had a legal right to was a concrete injury. The defendants were the CVS entities alone; the two companies that built and ran the contested inference were never parties. CVS settled individually and confidentially before any class ruling, so nothing about the technology was ever decided — and the revived statute has since produced more than twenty copycat suits, most alleging only a missing notice.

HireVue builds the video-assessment engine; hundreds of separate employers buy it, and every statutory duty that reaches the deployment lands on the buyer. This is the atlas's vendor-layer hiring case, and its record is unusual in three ways: it contains an input that was measured and then retired platform-wide with the arithmetic published, an audit whose narrow perimeter was publicized broadly and whose report is readable only under a nondisclosure agreement, and a mandated-transparency regime that inverted — the vendor holds the data and engages the auditor while the buyers hold the duty and post the result, so the same vendor-level numbers surfaced under four different employers' names. The only channel that ever attached a price was a private biometric-consent class action, and its settlement is preliminarily approved rather than paid, with no admission and no finding by any court or regulator.

A Deaf, Pawnee tax expert had worked five seasons for Intuit, been promoted once, led a team supporting about four hundred associates, earned a bonus every year and been encouraged to apply for the next role by a manager who sat on the hiring team. The employer, in other words, already held years of direct evidence about her. It routed her promotion through a roughly three-hour standardised assessment run on HireVue instead: about a dozen timed recorded video questions plus essay and multiple-choice sections, with instructions delivered audibly. She asked for human-generated real-time captioning. As alleged, she was told the platform's built-in subtitles could be enabled; when the assessment began there was no subtitle option, and she completed three hours of it on browser automatic captions she supplied herself. An automated rejection followed on 13 August 2024, then a feedback message advising her to be more concise, adapt her communication style, and practise active listening. On 19 March 2025 the ACLU, the ACLU of Colorado, Public Justice and Eisenberg & Baum filed a Complaint of Discrimination with the Colorado Civil Rights Division and the EEOC against both companies. Both call the allegations entirely without merit, and HireVue disputes that any AI assessment was used at all. No determination of any kind had surfaced by August 2026.

Per the EEOC's complaint, application software at the top of a tutoring company's hiring funnel was programmed to automatically reject female applicants aged 55 or older and male applicants aged 60 or older, turning away more than 200 qualified United-States applicants in a two-month window. Nothing was learned and nothing malfunctioned: the threshold was written by hand, computed from the birthdate field on the form, and fired before any person saw the application. It surfaced when one rejected applicant re-applied about a day later with an identical application and a more recent birthdate, and was offered an interview. The company paid $365,000 without admitting liability, and the decree that followed did not correct the rule — it prohibited collecting the field the rule read.

The chatbot that screens and schedules applicants for about nine in ten McDonald's franchises had quietly accumulated roughly 64 million application records — names, contact details, and every word applicants typed to the recruiter-bot — behind an admin password of 123456 and an API that never checked who was asking. Two outside researchers found both flaws in a cursory review, and the fix took thirty hours from a single disclosure email, with no regulator anywhere in the story. This is the atlas's custody case: the deployment where the decision path had a human check and the data path had none. What makes it a hiring case rather than a breach entry is what the store held, which is the employment-decision surface itself — who applied where, what they told a recruiting agent, the assessment step they were routed through, and their status history.

Civil-rights groups made Facebook strip age, gender and postal-code targeting from every US job ad in March 2019, in five coordinated settlements the company co-signed. Weeks later a peer-reviewed study bought paired ads with identical neutral targeting and measured them arriving at strongly gender- and race-skewed audiences — through the platform's own delivery stage, one layer below where the remedy had landed. The only delivery-layer control with numbers attached, a court-mandated variance controller with four-monthly reporting and an independent reviewer who recomputed the operator's figures to a 0.0 percent difference, is anchored to HOUSING ads; the employment extension is reported by Meta with no published metrics and no verification. The class charge that occupies that gap has been pending before the EEOC since December 2022, the court's oversight clock ran out in June 2026, and the people this screen reaches past never apply anywhere and appear in no record at all.

In 2020 pymetrics, a hiring-assessment vendor, paid a university team $104,465 — as a grant, delivered before the findings — to read the source code of the de-biasing gate inside its Soft Skills Platform, under a contract that let the auditors publish a negative result. They found the gate correctly implemented and their attempts to circumvent it defeated. The same contract had already put outside the examination the question of whether the games measure anything a job needs.

Sirius XM Radio licenses iCIMS, a commercial applicant-tracking system, and its AI candidate matching, shortlisting and sourcing features, and one employer stands alone as the defendant for what they are alleged to produce. The vendor is not sued. A Detroit information-technology professional pleads that he applied for about 150 positions in twelve months and was rejected for all but one, that the screening keys on data points proxying for race — educational institutions, employment history and postal codes — and that the platform penalises repeat applications, which he says he tried to defeat by applying from several email addresses. Read the register first, because nothing here has been decided: no court has ruled on the merits, on certification or on the fully briefed motion for judgment on the pleadings that has been under submission since March 2026, and the EEOC closed the charge with an express statement that it makes no determination. Every account of how the screen behaves is one side's pleading, and the employer denies liability.